Arcee AI, a U.S.-based open intelligence lab building high-efficiency, open-weight foundation models, announced the launch of Trinity Large, a 400-billion-parameter sparse Mixture-of-Experts (MoE) language model engineered for next-generation AI applications. Trinity Large delivers state-of-the-art performance with unmatched inference efficiency and multiple production-ready checkpoints for developers, enterprises, and researchers.
As part of the Trinity family of models, Trinity Large extends Arcee’s commitment to open-weight AI that scales from edge to cloud deployments a philosophy rooted in optimizing for performance per parameter and ownership of model weights.
“We like to say that we built Trinity so you can own it. Being able to say that about a frontier-level model is something we’re immeasurably proud of.”
Comprehensive Model Suite & Capabilities
Arcee is releasing three Trinity Large variants to support a broad spectrum of use cases:
-
Trinity-Large-Preview – A lightly post-trained, chat-ready instruct model designed for creative, narrative, and conversational workloads.
-
Trinity-Large-Base – A fully pre-trained checkpoint representing Arcee’s frontier-class foundation model after a complete 17T token training run.
-
TrueBase – An early pre-training checkpoint at 10T tokens with no instruct data or LR annealing, offering researchers a genuine baseline for foundational model behavior.
Trinity-Large features 256 experts with four active per token a high sparsity ratio that enables efficient computation while maintaining broad capability across math, reasoning, coding, and general knowledge tasks.
Also Read: GenAI Articul8 AI Series B Led by Adara, $500M+ Valuation
Performance & Efficiency
Arcee trained Trinity Large on 17 trillion tokens using 2048 NVIDIA B300 GPUs one of the largest such pre-training runs publicly disclosed for this hardware class. Optimized attention mechanisms and high sparsity enabled the team to achieve roughly 2–3× faster training and inference efficiency compared to peers in the same parameter class, without sacrificing model performance.
Data curation for Trinity Large was delivered in collaboration with DatologyAI, including a mix of programming, STEM, reasoning, multilingual content, and more than 8 trillion tokens of synthetic data generated to enhance performance across domains.
Preview Availability & Integration
Trinity-Large-Preview is now available through OpenRouter, free during the preview period through at least February 2026. Integrations with tooling platforms such as Kilo Code, Cline, and OpenCode are also supported at launch, enabling developers and enterprise users to rapidly embed the model into coding and agentic workflows.
“If you put this model into something real and it breaks, tell us. The fastest way for open models to get better is for people to actually use them, hard, in places that don’t look like benchmarks.”
Source: Arcee
